AI prices halve every six months.

Industrial robots took thirty years to drop 80% in price. From around $100,000 in 1995 to about $20,000 in 2025.

And that decline is slowing down.

Just process that gap for a second. Intelligence is getting cheap fast. Bodies are not. And two economists at Brookings just put out a paper arguing that this one asymmetry might decide everything about how AI actually plays out for the rest of us.

If you've been reading me for a while, you know I just wrote about the gentle singularity. About one guy building a $1.8 billion company with his brother. A data analyst curing his dog's cancer with ChatGPT. The whole thing.

That article was optimistic. This one is going to feel different.

Not because I changed my mind. Because I think it's a counter-argument to the singularity narrative that nobody on AI Twitter is actually engaging with.

The paper is called "Intelligence Saturation and the Future of Work." Konrad Kording and Ioana Marinescu. One neuroscientist, one economist, both at the University of Pennsylvania.

They decided to settle the fight. The one between AI people who think the singularity is at the door, and economists who think AI is just another tech wave.

And they came up with one phrase..

Intelligence saturation.

Here's the basic argument. The economy has two kinds of work.

Intelligence work. Sitting at a desk. Writing, analysing, coding, deciding. Anything you can do over Zoom.

And physical work. Building things, moving things, fixing things, cooking, driving, doing surgery, teaching kids in a real classroom.

These two are not the same. They don't scale the same way. And that turns out to be the whole story.

AI moves at computer-science speeds. Compute doubles every two years. AI inference cost halves every six months. Energy efficiency for compute improves about a hundred times every decade.

The physical world doesn't move like that at all.

Robot prices took thirty years to drop 80%. AI prices fall that much in a year and a half. The pace of physical-world improvement is on a different planet from the pace of intelligence improvement.

And the interesting part is..

You need both. You can't have one without the other. They're complements, not substitutes.

Take education. The intelligence part is the curriculum, the lesson plan, the textbook. AI can make all of that better, faster, cheaper. Cool.

But somebody still has to be in the classroom. Somebody still has to manage a room of kids, run the activity, hold the energy together.

During COVID, when learning went virtual, kids fell behind. Massively. The intelligence inputs got better. The physical part collapsed. And outcomes declined.

Same for surgery. Same for a restaurant. Same for fixing the toilet in your bathroom.

The authors says.. With perfect intelligence, you can use a body at its physical limit. Beyond that point, more intelligence yields nothing. The body is the bottleneck. Not the brain.

That's intelligence saturation.

And once you start looking for it, you see it everywhere.

The clearest example for me was the Denmark study.

Researchers tracked what happened when companies actually rolled out AI chatbots at the firm level. Real workers. Real jobs. Real adoption.

You know what they found?

Zero effect on employment. And the workers who used AI chatbots saved about 3% of their time on average.

3%.

A US study covering the whole working-age population came out with a similar number. 1.4% productivity gain across the population.

Not 14%. Not 40%. One point four percent.

The way we talk about AI on Twitter, in podcasts, in keynote decks.. it sounds like the economy is being rewritten in real time. The actual data on the ground is.. people saving 3% of their time.

That's not nothing. But it's also not the singularity.

And this isn't the first time we've seen this pattern. The previous big tech wave, the entire ICT revolution.. computers, internet, mobile, software, all of it.. contributed at most one percentage point to annual GDP growth at its peak. Decades of compounding tech, and the ceiling was one point.

That gap between what tech promises and what shows up in the actual economy. That's what the authors are calling intelligence saturation.

Viktor is an AI coworker that lives in Slack, right where your team already works.

Message Viktor like a teammate: "pull last quarter's revenue by channel," or "build a dashboard for our board meeting."

Viktor connects to your tools, does the work, and delivers the actual report, spreadsheet, or dashboard. Not a summary. The real thing.

There’s no new software to adopt and no one to train.

Most teams start with one task. Within a week, Viktor is handling half of their ops.

Add Viktor to Slack for free.

They ran simulations of wages over time as AI keeps automating intelligence tasks.

And the curve is shaped like a hump.

Wages go up first. They peak. Then they start falling.

Here's why. When AI starts automating intelligence work, the humans still doing intelligence work get more AI capital per person. They become more productive. Wages go up. Companies pay more for the humans who are still in the room.

But as automation keeps going, more people get pushed out of the intelligence sector entirely. They have to go somewhere. The only place left is the physical sector.

And when you crowd a bunch of workers into a sector that hasn't grown to receive them.. wages there fall.

At some point, that crowding effect catches up to the productivity effect. And total wages start declining.

The paper estimates we're already maybe 14% into the automation of intelligence tasks, just from ICT and earlier digital tech. That number is climbing fast now because of LLMs.

So the ride looks like this. AI makes us richer. Then it makes us poorer. Then if AI keeps growing past full automation, wages stabilise and start growing again, but only because more AI keeps lifting productivity for the workers stuck doing physical work.

The shape of the curve depends on parameters the authors can't measure precisely. But the hump itself shows up across most of their simulations. The exact peak moves. The exact drop moves. The pattern doesn't.

Oh, and one more thing.

"Physical" in this paper doesn't mean "blue collar." A surgeon is physical. A live courtroom lawyer is physical. A classroom teacher is physical. A nurse, a chef, an electrician, a contractor, anyone whose value depends on being there in person.

About 30% of US workers are in clearly physical occupations today. That number goes up as more intelligence work gets automated. And the wages in that growing physical sector go down because of the crowding.

The Sam Altman says intelligence is the limiting input on growth. Solve intelligence and everything else unlocks. Energy. Biology. Manufacturing. Governance. All of it falls.

This paper is saying something different. It's saying intelligence is one input. The body of the economy is another. And no matter how cheap intelligence gets, the body sets the ceiling.

A car needs steel. A meal needs ingredients and a kitchen. Smarter recipes don't make food taste dramatically better past a point. Smarter controllers don't bypass the physical speed of a factory line.

I think the AI side has been a little drunk on the idea that intelligence alone changes everything. And I've been part of that conversation, honestly.

But I also don't fully buy the economists' framing. Because their model assumes the physical sector stays roughly fixed in capability. And that's not obviously true. Robots are getting better. AI agents are starting to get embodied. The boundary between virtual work and physical work might move.

The authors actually address this. They say if technological progress turns physical tasks into intelligence ones over time, the substitutability between the two sectors goes up. The ceiling rises. Saturation goes away.

That feels like the actual question. Not whether intelligence has a ceiling. It does.

The question is whether we can keep moving the ceiling.

From my point of view, that's the real fight of the next decade. Not AGI vs not-AGI. But how fast can we take physical bottlenecks like construction, manufacturing, healthcare delivery, last-mile logistics, energy generation.. and turn them into things AI and robots can handle without humans being the rate-limiter.

If we move fast on that, the gentle singularity keeps going.

If we don't, wages do that hump. They rise. They peak. They fall.

And the people who get stuck on the falling side of the curve are not the people writing papers about it.

The paper has a policy section. Slow down automation. Invest in physical capital. Wage insurance for displaced workers. Reasonable ideas. Almost certainly not happening at the scale needed in the US given everything else going on right now.

So we're left with what the paper actually says.

Wages go up first. Then they fall. Then maybe they go back up, but only if AI keeps growing and the physical bottleneck starts to soften.

Right now, everyone is talking about the up part. Nobody is talking about the part that comes after.

But it's there.

I'm not saying this to scare you.

You don't get the miracles without also getting the dislocations. The same forces that let one guy build a $1.8 billion company from his living room are the forces that push a lot of other people into a smaller and smaller corner of the economy.

That's not a doom take. That's just what the model says.

What do you think? Is the physical world actually a hard ceiling on AI's economic impact, or do you think robotics and embodiment will break through it sooner than economists expect?

Let me know. I read everything.

If you made it this far, you're not a casual reader. You actually think about this stuff.

So here's my ask. If this article made you think, even a little, share it with one person. Just one. Someone who's in the AI space. Someone who reads. Someone who would actually sit with these ideas instead of scrolling past them.

That's how this newsletter grows. Not through ads or algorithms. Through you sending it to someone and saying "read this."

Research Paper: Intelligence saturation and the future of work